Add within_gradient (#434)

* add within_gradient

* add ForwardDiff method

* docs

* use in softmax too

Author: mcabbott

Project.toml

@@ -22,6 +22,7 @@ CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
 ChainRulesTestUtils = "cdddcdb0-9152-4a09-a978-84456f9df70a"
 Documenter = "e30172f5-a6a5-5a46-863b-614d45cd2de4"
 FiniteDifferences = "26cc04aa-876d-5657-8c51-4c34ba976000"
+ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
 Logging = "56ddb016-857b-54e1-b83d-db4d58db5568"
 NNlibCUDA = "a00861dc-f156-4864-bf3c-e6376f28a68d"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
@@ -31,4 +32,4 @@ UnicodePlots = "b8865327-cd53-5732-bb35-84acbb429228"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [targets]
-test = ["ChainRulesTestUtils", "CUDA", "Documenter", "FiniteDifferences", "Logging", "NNlibCUDA", "Random", "StableRNGs", "Test", "UnicodePlots", "Zygote"]
+test = ["ChainRulesTestUtils", "CUDA", "Documenter", "FiniteDifferences", "ForwardDiff", "Logging", "NNlibCUDA", "Random", "StableRNGs", "Test", "UnicodePlots", "Zygote"]

docs/src/reference.md

@@ -132,4 +132,5 @@ ctc_loss
 ```@docs
 logsumexp
 NNlib.glu
+NNlib.within_gradient
 ```

src/NNlib.jl

@@ -87,6 +87,12 @@ export upsample_nearest, ∇upsample_nearest,
 include("gather.jl")
 include("scatter.jl")
 include("utils.jl")
+@init @require ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210" begin
+    using .ForwardDiff
+    within_gradient(x::ForwardDiff.Dual) = true
+    within_gradient(x::AbstractArray{<:ForwardDiff.Dual}) = true
+end
+
 include("sampling.jl")
 include("functions.jl")
 

src/softmax.jl

@@ -69,7 +69,7 @@ function softmax!(out::AbstractArray{T}, x::AbstractArray; dims = 1) where {T}
 end
 
 function ∇softmax_data(dy::AbstractArray{T}, y::AbstractArray{S}; dims = 1) where {T,S}
-    dx = if within_grad()
+    dx = if within_gradient(y)
         tmp = dy .* y
         tmp .- y .* sum(tmp; dims)
     else
@@ -88,9 +88,6 @@ function rrule(::typeof(softmax), x; dims = 1)
     return y, softmax_pullback
 end
 
-within_grad() = false
-rrule(::typeof(within_grad)) = true, _ -> (NoTangent(),)
-
 fast_maximum(x::AbstractArray{T}; dims) where {T} = @fastmath reduce(max, x; dims, init = float(T)(-Inf))
 
 """

src/utils.jl

@@ -1,3 +1,55 @@
+"""
+    within_gradient(x) --> Bool
+
+Returns `false` except when used inside a `gradient` call, when it returns `true`.
+Useful for Flux regularisation layers which behave differently during training and inference.
+
+This should work with any ChainRules-based differentiation package, in which case `x` is ignored.
+But Tracker.jl overloads `with_gradient(x::TrackedArray)`, thus for widest use you should
+pass it an array whose gradient is of interest.
+There is also an overload for ForwardDiff.jl's `Dual` types (and arrays of them).
+
+# Examples
+```
+julia> using ForwardDiff, Zygote, NNlib
+
+julia> f_good(x) = if NNlib.within_gradient(x)
+                     @show 10x
+                   else
+                     x
+                   end;
+
+julia> Zygote.withgradient(f_good, 1.0)
+10x = 10.0
+(val = 10.0, grad = (10.0,))
+
+julia> ForwardDiff.derivative(f_good, 1.0)
+10x = Dual{ForwardDiff.Tag{typeof(f_good), Float64}}(10.0,10.0)
+10.0
+
+julia> f_bad(x, y) = if any(NNlib.within_gradient, (x, y))
+                       @show x * y
+                     else
+                       x / y
+                     end;
+
+julia> Zygote.withgradient(f_bad, 2.0, 3.0)
+(val = 0.6666666666666666, grad = (0.3333333333333333, -0.2222222222222222))
+
+julia> ForwardDiff.derivative(x -> f_bad(x, 3.0), 2.0)
+x * y = Dual{ForwardDiff.Tag{var"#9#10", Float64}}(6.0,3.0)
+3.0
+```
+
+What goes wrong in `f_bad` is that Zygote knows `any` to be non-differentiable,
+and thus completely ignores its contents. This is not a perfect mechanism,
+and the only style recommended is precisely that of `f_good` above.
+"""
+within_gradient(x) = false
+
+ChainRulesCore.rrule(::typeof(within_gradient), x) = true, _ -> (NoTangent(), NoTangent())
+
+
 """
     safe_div(x, y)
 

test/runtests.jl

@@ -2,6 +2,7 @@ using NNlib, Test, Statistics, Random
 using ChainRulesCore, ChainRulesTestUtils
 using Base.Broadcast: broadcasted
 import FiniteDifferences
+import ForwardDiff
 import Zygote
 using Zygote: gradient
 using StableRNGs

test/utils.jl

@@ -1,3 +1,9 @@
+@testset "within_gradient" begin
+    @test NNlib.within_gradient([1.0]) === false
+    @test gradient(x -> NNlib.within_gradient(x) * x, 2.0) == (1.0,)
+    @test NNlib.within_gradient([ForwardDiff.Dual(1.0, 2)]) === true
+end
+
 @testset "maximum_dims" begin
     ind1 = [1,2,3,4,5,6]
     @test NNlib.maximum_dims(ind1) == (6,)